kernel/bpf/bpf_local_storage.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /* Copyright (c) 2019 Facebook  */
   3 #include <linux/rculist.h>
   4 #include <linux/list.h>
   5 #include <linux/hash.h>
   6 #include <linux/types.h>
   7 #include <linux/spinlock.h>
   8 #include <linux/bpf.h>
   9 #include <linux/btf_ids.h>
  10 #include <linux/bpf_local_storage.h>
  11 #include <net/sock.h>
  12 #include <uapi/linux/sock_diag.h>
  13 #include <uapi/linux/btf.h>
  14
  15 #define BPF_LOCAL_STORAGE_CREATE_FLAG_MASK (BPF_F_NO_PREALLOC | BPF_F_CLONE)
  16
  17 static struct bpf_local_storage_map_bucket *
  18 select_bucket(struct bpf_local_storage_map *smap,
  19               struct bpf_local_storage_elem *selem)
  20 {
  21         return &smap->buckets[hash_ptr(selem, smap->bucket_log)];
  22 }
  23
  24 static int mem_charge(struct bpf_local_storage_map *smap, void *owner, u32 size)
  25 {
  26         struct bpf_map *map = &smap->map;
  27
  28         if (!map->ops->map_local_storage_charge)
  29                 return 0;
  30
  31         return map->ops->map_local_storage_charge(smap, owner, size);
  32 }
  33
  34 static void mem_uncharge(struct bpf_local_storage_map *smap, void *owner,
  35                          u32 size)
  36 {
  37         struct bpf_map *map = &smap->map;
  38
  39         if (map->ops->map_local_storage_uncharge)
  40                 map->ops->map_local_storage_uncharge(smap, owner, size);
  41 }
  42
  43 static struct bpf_local_storage __rcu **
  44 owner_storage(struct bpf_local_storage_map *smap, void *owner)
  45 {
  46         struct bpf_map *map = &smap->map;
  47
  48         return map->ops->map_owner_storage_ptr(owner);
  49 }
  50
  51 static bool selem_linked_to_storage_lockless(const struct bpf_local_storage_elem *selem)
  52 {
  53         return !hlist_unhashed_lockless(&selem->snode);
  54 }
  55
  56 static bool selem_linked_to_storage(const struct bpf_local_storage_elem *selem)
  57 {
  58         return !hlist_unhashed(&selem->snode);
  59 }
  60
  61 static bool selem_linked_to_map_lockless(const struct bpf_local_storage_elem *selem)
  62 {
  63         return !hlist_unhashed_lockless(&selem->map_node);
  64 }
  65
  66 static bool selem_linked_to_map(const struct bpf_local_storage_elem *selem)
  67 {
  68         return !hlist_unhashed(&selem->map_node);
  69 }
  70
  71 struct bpf_local_storage_elem *
  72 bpf_selem_alloc(struct bpf_local_storage_map *smap, void *owner,
  73                 void *value, bool charge_mem)
  74 {
  75         struct bpf_local_storage_elem *selem;
  76
  77         if (charge_mem && mem_charge(smap, owner, smap->elem_size))
  78                 return NULL;
  79
  80         selem = kzalloc(smap->elem_size, GFP_ATOMIC | __GFP_NOWARN);
  81         if (selem) {
  82                 if (value)
  83                         copy_map_value(&smap->map, SDATA(selem)->data, value);
  84                 return selem;
  85         }
  86
  87         if (charge_mem)
  88                 mem_uncharge(smap, owner, smap->elem_size);
  89
  90         return NULL;
  91 }
  92
  93 /* local_storage->lock must be held and selem->local_storage == local_storage.
  94  * The caller must ensure selem->smap is still valid to be
  95  * dereferenced for its smap->elem_size and smap->cache_idx.
  96  */
  97 bool bpf_selem_unlink_storage_nolock(struct bpf_local_storage *local_storage,
  98                                      struct bpf_local_storage_elem *selem,
  99                                      bool uncharge_mem)
 100 {
 101         struct bpf_local_storage_map *smap;
 102         bool free_local_storage;
 103         void *owner;
 104
 105         smap = rcu_dereference(SDATA(selem)->smap);
 106         owner = local_storage->owner;
 107
 108         /* All uncharging on the owner must be done first.
 109          * The owner may be freed once the last selem is unlinked
 110          * from local_storage.
 111          */
 112         if (uncharge_mem)
 113                 mem_uncharge(smap, owner, smap->elem_size);
 114
 115         free_local_storage = hlist_is_singular_node(&selem->snode,
 116                                                     &local_storage->list);
 117         if (free_local_storage) {
 118                 mem_uncharge(smap, owner, sizeof(struct bpf_local_storage));
 119                 local_storage->owner = NULL;
 120
 121                 /* After this RCU_INIT, owner may be freed and cannot be used */
 122                 RCU_INIT_POINTER(*owner_storage(smap, owner), NULL);
 123
 124                 /* local_storage is not freed now.  local_storage->lock is
 125                  * still held and raw_spin_unlock_bh(&local_storage->lock)
 126                  * will be done by the caller.
 127                  *
 128                  * Although the unlock will be done under
 129                  * rcu_read_lock(),  it is more intutivie to
 130                  * read if kfree_rcu(local_storage, rcu) is done
 131                  * after the raw_spin_unlock_bh(&local_storage->lock).
 132                  *
 133                  * Hence, a "bool free_local_storage" is returned
 134                  * to the caller which then calls the kfree_rcu()
 135                  * after unlock.
 136                  */
 137         }
 138         hlist_del_init_rcu(&selem->snode);
 139         if (rcu_access_pointer(local_storage->cache[smap->cache_idx]) ==
 140             SDATA(selem))
 141                 RCU_INIT_POINTER(local_storage->cache[smap->cache_idx], NULL);
 142
 143         kfree_rcu(selem, rcu);
 144
 145         return free_local_storage;
 146 }
 147
 148 static void __bpf_selem_unlink_storage(struct bpf_local_storage_elem *selem)
 149 {
 150         struct bpf_local_storage *local_storage;
 151         bool free_local_storage = false;
 152
 153         if (unlikely(!selem_linked_to_storage_lockless(selem)))
 154                 /* selem has already been unlinked from sk */
 155                 return;
 156
 157         local_storage = rcu_dereference(selem->local_storage);
 158         raw_spin_lock_bh(&local_storage->lock);
 159         if (likely(selem_linked_to_storage(selem)))
 160                 free_local_storage = bpf_selem_unlink_storage_nolock(
 161                         local_storage, selem, true);
 162         raw_spin_unlock_bh(&local_storage->lock);
 163
 164         if (free_local_storage)
 165                 kfree_rcu(local_storage, rcu);
 166 }
 167
 168 void bpf_selem_link_storage_nolock(struct bpf_local_storage *local_storage,
 169                                    struct bpf_local_storage_elem *selem)
 170 {
 171         RCU_INIT_POINTER(selem->local_storage, local_storage);
 172         hlist_add_head_rcu(&selem->snode, &local_storage->list);
 173 }
 174
 175 void bpf_selem_unlink_map(struct bpf_local_storage_elem *selem)
 176 {
 177         struct bpf_local_storage_map *smap;
 178         struct bpf_local_storage_map_bucket *b;
 179
 180         if (unlikely(!selem_linked_to_map_lockless(selem)))
 181                 /* selem has already be unlinked from smap */
 182                 return;
 183
 184         smap = rcu_dereference(SDATA(selem)->smap);
 185         b = select_bucket(smap, selem);
 186         raw_spin_lock_bh(&b->lock);
 187         if (likely(selem_linked_to_map(selem)))
 188                 hlist_del_init_rcu(&selem->map_node);
 189         raw_spin_unlock_bh(&b->lock);
 190 }
 191
 192 void bpf_selem_link_map(struct bpf_local_storage_map *smap,
 193                         struct bpf_local_storage_elem *selem)
 194 {
 195         struct bpf_local_storage_map_bucket *b = select_bucket(smap, selem);
 196
 197         raw_spin_lock_bh(&b->lock);
 198         RCU_INIT_POINTER(SDATA(selem)->smap, smap);
 199         hlist_add_head_rcu(&selem->map_node, &b->list);
 200         raw_spin_unlock_bh(&b->lock);
 201 }
 202
 203 void bpf_selem_unlink(struct bpf_local_storage_elem *selem)
 204 {
 205         /* Always unlink from map before unlinking from local_storage
 206          * because selem will be freed after successfully unlinked from
 207          * the local_storage.
 208          */
 209         bpf_selem_unlink_map(selem);
 210         __bpf_selem_unlink_storage(selem);
 211 }
 212
 213 struct bpf_local_storage_data *
 214 bpf_local_storage_lookup(struct bpf_local_storage *local_storage,
 215                          struct bpf_local_storage_map *smap,
 216                          bool cacheit_lockit)
 217 {
 218         struct bpf_local_storage_data *sdata;
 219         struct bpf_local_storage_elem *selem;
 220
 221         /* Fast path (cache hit) */
 222         sdata = rcu_dereference(local_storage->cache[smap->cache_idx]);
 223         if (sdata && rcu_access_pointer(sdata->smap) == smap)
 224                 return sdata;
 225
 226         /* Slow path (cache miss) */
 227         hlist_for_each_entry_rcu(selem, &local_storage->list, snode)
 228                 if (rcu_access_pointer(SDATA(selem)->smap) == smap)
 229                         break;
 230
 231         if (!selem)
 232                 return NULL;
 233
 234         sdata = SDATA(selem);
 235         if (cacheit_lockit) {
 236                 /* spinlock is needed to avoid racing with the
 237                  * parallel delete.  Otherwise, publishing an already
 238                  * deleted sdata to the cache will become a use-after-free
 239                  * problem in the next bpf_local_storage_lookup().
 240                  */
 241                 raw_spin_lock_bh(&local_storage->lock);
 242                 if (selem_linked_to_storage(selem))
 243                         rcu_assign_pointer(local_storage->cache[smap->cache_idx],
 244                                            sdata);
 245                 raw_spin_unlock_bh(&local_storage->lock);
 246         }
 247
 248         return sdata;
 249 }
 250
 251 static int check_flags(const struct bpf_local_storage_data *old_sdata,
 252                        u64 map_flags)
 253 {
 254         if (old_sdata && (map_flags & ~BPF_F_LOCK) == BPF_NOEXIST)
 255                 /* elem already exists */
 256                 return -EEXIST;
 257
 258         if (!old_sdata && (map_flags & ~BPF_F_LOCK) == BPF_EXIST)
 259                 /* elem doesn't exist, cannot update it */
 260                 return -ENOENT;
 261
 262         return 0;
 263 }
 264
 265 int bpf_local_storage_alloc(void *owner,
 266                             struct bpf_local_storage_map *smap,
 267                             struct bpf_local_storage_elem *first_selem)
 268 {
 269         struct bpf_local_storage *prev_storage, *storage;
 270         struct bpf_local_storage **owner_storage_ptr;
 271         int err;
 272
 273         err = mem_charge(smap, owner, sizeof(*storage));
 274         if (err)
 275                 return err;
 276
 277         storage = kzalloc(sizeof(*storage), GFP_ATOMIC | __GFP_NOWARN);
 278         if (!storage) {
 279                 err = -ENOMEM;
 280                 goto uncharge;
 281         }
 282
 283         INIT_HLIST_HEAD(&storage->list);
 284         raw_spin_lock_init(&storage->lock);
 285         storage->owner = owner;
 286
 287         bpf_selem_link_storage_nolock(storage, first_selem);
 288         bpf_selem_link_map(smap, first_selem);
 289
 290         owner_storage_ptr =
 291                 (struct bpf_local_storage **)owner_storage(smap, owner);
 292         /* Publish storage to the owner.
 293          * Instead of using any lock of the kernel object (i.e. owner),
 294          * cmpxchg will work with any kernel object regardless what
 295          * the running context is, bh, irq...etc.
 296          *
 297          * From now on, the owner->storage pointer (e.g. sk->sk_bpf_storage)
 298          * is protected by the storage->lock.  Hence, when freeing
 299          * the owner->storage, the storage->lock must be held before
 300          * setting owner->storage ptr to NULL.
 301          */
 302         prev_storage = cmpxchg(owner_storage_ptr, NULL, storage);
 303         if (unlikely(prev_storage)) {
 304                 bpf_selem_unlink_map(first_selem);
 305                 err = -EAGAIN;
 306                 goto uncharge;
 307
 308                 /* Note that even first_selem was linked to smap's
 309                  * bucket->list, first_selem can be freed immediately
 310                  * (instead of kfree_rcu) because
 311                  * bpf_local_storage_map_free() does a
 312                  * synchronize_rcu() before walking the bucket->list.
 313                  * Hence, no one is accessing selem from the
 314                  * bucket->list under rcu_read_lock().
 315                  */
 316         }
 317
 318         return 0;
 319
 320 uncharge:
 321         kfree(storage);
 322         mem_uncharge(smap, owner, sizeof(*storage));
 323         return err;
 324 }
 325
 326 /* sk cannot be going away because it is linking new elem
 327  * to sk->sk_bpf_storage. (i.e. sk->sk_refcnt cannot be 0).
 328  * Otherwise, it will become a leak (and other memory issues
 329  * during map destruction).
 330  */
 331 struct bpf_local_storage_data *
 332 bpf_local_storage_update(void *owner, struct bpf_local_storage_map *smap,
 333                          void *value, u64 map_flags)
 334 {
 335         struct bpf_local_storage_data *old_sdata = NULL;
 336         struct bpf_local_storage_elem *selem;
 337         struct bpf_local_storage *local_storage;
 338         int err;
 339
 340         /* BPF_EXIST and BPF_NOEXIST cannot be both set */
 341         if (unlikely((map_flags & ~BPF_F_LOCK) > BPF_EXIST) ||
 342             /* BPF_F_LOCK can only be used in a value with spin_lock */
 343             unlikely((map_flags & BPF_F_LOCK) &&
 344                      !map_value_has_spin_lock(&smap->map)))
 345                 return ERR_PTR(-EINVAL);
 346
 347         local_storage = rcu_dereference(*owner_storage(smap, owner));
 348         if (!local_storage || hlist_empty(&local_storage->list)) {
 349                 /* Very first elem for the owner */
 350                 err = check_flags(NULL, map_flags);
 351                 if (err)
 352                         return ERR_PTR(err);
 353
 354                 selem = bpf_selem_alloc(smap, owner, value, true);
 355                 if (!selem)
 356                         return ERR_PTR(-ENOMEM);
 357
 358                 err = bpf_local_storage_alloc(owner, smap, selem);
 359                 if (err) {
 360                         kfree(selem);
 361                         mem_uncharge(smap, owner, smap->elem_size);
 362                         return ERR_PTR(err);
 363                 }
 364
 365                 return SDATA(selem);
 366         }
 367
 368         if ((map_flags & BPF_F_LOCK) && !(map_flags & BPF_NOEXIST)) {
 369                 /* Hoping to find an old_sdata to do inline update
 370                  * such that it can avoid taking the local_storage->lock
 371                  * and changing the lists.
 372                  */
 373                 old_sdata =
 374                         bpf_local_storage_lookup(local_storage, smap, false);
 375                 err = check_flags(old_sdata, map_flags);
 376                 if (err)
 377                         return ERR_PTR(err);
 378                 if (old_sdata && selem_linked_to_storage_lockless(SELEM(old_sdata))) {
 379                         copy_map_value_locked(&smap->map, old_sdata->data,
 380                                               value, false);
 381                         return old_sdata;
 382                 }
 383         }
 384
 385         raw_spin_lock_bh(&local_storage->lock);
 386
 387         /* Recheck local_storage->list under local_storage->lock */
 388         if (unlikely(hlist_empty(&local_storage->list))) {
 389                 /* A parallel del is happening and local_storage is going
 390                  * away.  It has just been checked before, so very
 391                  * unlikely.  Return instead of retry to keep things
 392                  * simple.
 393                  */
 394                 err = -EAGAIN;
 395                 goto unlock_err;
 396         }
 397
 398         old_sdata = bpf_local_storage_lookup(local_storage, smap, false);
 399         err = check_flags(old_sdata, map_flags);
 400         if (err)
 401                 goto unlock_err;
 402
 403         if (old_sdata && (map_flags & BPF_F_LOCK)) {
 404                 copy_map_value_locked(&smap->map, old_sdata->data, value,
 405                                       false);
 406                 selem = SELEM(old_sdata);
 407                 goto unlock;
 408         }
 409
 410         /* local_storage->lock is held.  Hence, we are sure
 411          * we can unlink and uncharge the old_sdata successfully
 412          * later.  Hence, instead of charging the new selem now
 413          * and then uncharge the old selem later (which may cause
 414          * a potential but unnecessary charge failure),  avoid taking
 415          * a charge at all here (the "!old_sdata" check) and the
 416          * old_sdata will not be uncharged later during
 417          * bpf_selem_unlink_storage_nolock().
 418          */
 419         selem = bpf_selem_alloc(smap, owner, value, !old_sdata);
 420         if (!selem) {
 421                 err = -ENOMEM;
 422                 goto unlock_err;
 423         }
 424
 425         /* First, link the new selem to the map */
 426         bpf_selem_link_map(smap, selem);
 427
 428         /* Second, link (and publish) the new selem to local_storage */
 429         bpf_selem_link_storage_nolock(local_storage, selem);
 430
 431         /* Third, remove old selem, SELEM(old_sdata) */
 432         if (old_sdata) {
 433                 bpf_selem_unlink_map(SELEM(old_sdata));
 434                 bpf_selem_unlink_storage_nolock(local_storage, SELEM(old_sdata),
 435                                                 false);
 436         }
 437
 438 unlock:
 439         raw_spin_unlock_bh(&local_storage->lock);
 440         return SDATA(selem);
 441
 442 unlock_err:
 443         raw_spin_unlock_bh(&local_storage->lock);
 444         return ERR_PTR(err);
 445 }
 446
 447 u16 bpf_local_storage_cache_idx_get(struct bpf_local_storage_cache *cache)
 448 {
 449         u64 min_usage = U64_MAX;
 450         u16 i, res = 0;
 451
 452         spin_lock(&cache->idx_lock);
 453
 454         for (i = 0; i < BPF_LOCAL_STORAGE_CACHE_SIZE; i++) {
 455                 if (cache->idx_usage_counts[i] < min_usage) {
 456                         min_usage = cache->idx_usage_counts[i];
 457                         res = i;
 458
 459                         /* Found a free cache_idx */
 460                         if (!min_usage)
 461                                 break;
 462                 }
 463         }
 464         cache->idx_usage_counts[res]++;
 465
 466         spin_unlock(&cache->idx_lock);
 467
 468         return res;
 469 }
 470
 471 void bpf_local_storage_cache_idx_free(struct bpf_local_storage_cache *cache,
 472                                       u16 idx)
 473 {
 474         spin_lock(&cache->idx_lock);
 475         cache->idx_usage_counts[idx]--;
 476         spin_unlock(&cache->idx_lock);
 477 }
 478
 479 void bpf_local_storage_map_free(struct bpf_local_storage_map *smap)
 480 {
 481         struct bpf_local_storage_elem *selem;
 482         struct bpf_local_storage_map_bucket *b;
 483         unsigned int i;
 484
 485         /* Note that this map might be concurrently cloned from
 486          * bpf_sk_storage_clone. Wait for any existing bpf_sk_storage_clone
 487          * RCU read section to finish before proceeding. New RCU
 488          * read sections should be prevented via bpf_map_inc_not_zero.
 489          */
 490         synchronize_rcu();
 491
 492         /* bpf prog and the userspace can no longer access this map
 493          * now.  No new selem (of this map) can be added
 494          * to the owner->storage or to the map bucket's list.
 495          *
 496          * The elem of this map can be cleaned up here
 497          * or when the storage is freed e.g.
 498          * by bpf_sk_storage_free() during __sk_destruct().
 499          */
 500         for (i = 0; i < (1U << smap->bucket_log); i++) {
 501                 b = &smap->buckets[i];
 502
 503                 rcu_read_lock();
 504                 /* No one is adding to b->list now */
 505                 while ((selem = hlist_entry_safe(
 506                                 rcu_dereference_raw(hlist_first_rcu(&b->list)),
 507                                 struct bpf_local_storage_elem, map_node))) {
 508                         bpf_selem_unlink(selem);
 509                         cond_resched_rcu();
 510                 }
 511                 rcu_read_unlock();
 512         }
 513
 514         /* While freeing the storage we may still need to access the map.
 515          *
 516          * e.g. when bpf_sk_storage_free() has unlinked selem from the map
 517          * which then made the above while((selem = ...)) loop
 518          * exit immediately.
 519          *
 520          * However, while freeing the storage one still needs to access the
 521          * smap->elem_size to do the uncharging in
 522          * bpf_selem_unlink_storage_nolock().
 523          *
 524          * Hence, wait another rcu grace period for the storage to be freed.
 525          */
 526         synchronize_rcu();
 527
 528         kvfree(smap->buckets);
 529         kfree(smap);
 530 }
 531
 532 int bpf_local_storage_map_alloc_check(union bpf_attr *attr)
 533 {
 534         if (attr->map_flags & ~BPF_LOCAL_STORAGE_CREATE_FLAG_MASK ||
 535             !(attr->map_flags & BPF_F_NO_PREALLOC) ||
 536             attr->max_entries ||
 537             attr->key_size != sizeof(int) || !attr->value_size ||
 538             /* Enforce BTF for userspace sk dumping */
 539             !attr->btf_key_type_id || !attr->btf_value_type_id)
 540                 return -EINVAL;
 541
 542         if (!bpf_capable())
 543                 return -EPERM;
 544
 545         if (attr->value_size > BPF_LOCAL_STORAGE_MAX_VALUE_SIZE)
 546                 return -E2BIG;
 547
 548         return 0;
 549 }
 550
 551 struct bpf_local_storage_map *bpf_local_storage_map_alloc(union bpf_attr *attr)
 552 {
 553         struct bpf_local_storage_map *smap;
 554         unsigned int i;
 555         u32 nbuckets;
 556         u64 cost;
 557         int ret;
 558
 559         smap = kzalloc(sizeof(*smap), GFP_USER | __GFP_NOWARN);
 560         if (!smap)
 561                 return ERR_PTR(-ENOMEM);
 562         bpf_map_init_from_attr(&smap->map, attr);
 563
 564         nbuckets = roundup_pow_of_two(num_possible_cpus());
 565         /* Use at least 2 buckets, select_bucket() is undefined behavior with 1 bucket */
 566         nbuckets = max_t(u32, 2, nbuckets);
 567         smap->bucket_log = ilog2(nbuckets);
 568         cost = sizeof(*smap->buckets) * nbuckets + sizeof(*smap);
 569
 570         ret = bpf_map_charge_init(&smap->map.memory, cost);
 571         if (ret < 0) {
 572                 kfree(smap);
 573                 return ERR_PTR(ret);
 574         }
 575
 576         smap->buckets = kvcalloc(sizeof(*smap->buckets), nbuckets,
 577                                  GFP_USER | __GFP_NOWARN);
 578         if (!smap->buckets) {
 579                 bpf_map_charge_finish(&smap->map.memory);
 580                 kfree(smap);
 581                 return ERR_PTR(-ENOMEM);
 582         }
 583
 584         for (i = 0; i < nbuckets; i++) {
 585                 INIT_HLIST_HEAD(&smap->buckets[i].list);
 586                 raw_spin_lock_init(&smap->buckets[i].lock);
 587         }
 588
 589         smap->elem_size =
 590                 sizeof(struct bpf_local_storage_elem) + attr->value_size;
 591
 592         return smap;
 593 }
 594
 595 int bpf_local_storage_map_check_btf(const struct bpf_map *map,
 596                                     const struct btf *btf,
 597                                     const struct btf_type *key_type,
 598                                     const struct btf_type *value_type)
 599 {
 600         u32 int_data;
 601
 602         if (BTF_INFO_KIND(key_type->info) != BTF_KIND_INT)
 603                 return -EINVAL;
 604
 605         int_data = *(u32 *)(key_type + 1);
 606         if (BTF_INT_BITS(int_data) != 32 || BTF_INT_OFFSET(int_data))
 607                 return -EINVAL;
 608
 609         return 0;
 610 }